AgentSkillsCN

prep-spec

为实施准备一份规范,邀请四位代理人共同审阅。每位代理人将带来不同的视角。

SKILL.md
--- frontmatter
name: prep-spec
description: Prepare a spec for implementation with 4-agent review. Each agent brings a different perspective.
when_to_use: before implementing any feature spec, when spec status is not "prepped"
version: 3.0.0

Prep Spec

Four-perspective spec review. Each agent thinks differently — no checklists, just perspectives.

Announce at start: "I'm using the prep-spec skill to prepare this spec for implementation."

Quick Start

code
/prep-spec features/p104_feature.md

The Four Perspectives

code
┌─────────────────────────────────────────────────────┐
│               Four Perspectives                     │
├─────────────────┬─────────────────┬─────────────────┤
│       UX        │    Architect    │   Lean Coach    │
│   (the user)    │  (production)   │   (simplest?)   │
├─────────────────┴─────────────────┴─────────────────┤
│                    Alignment                        │
│         (strategy + consistency check)              │
└─────────────────────────────────────────────────────┘
AgentKey QuestionWhat it catches
UX"How does this feel to use?"User friction, edge cases, accessibility
Architect"Will we regret this in 6 months?"Tech debt, failure modes, pattern violations
Lean Coach"Can we build less and still learn?"Scope creep, business misalignment, wasted effort
Alignment"Does this fit our strategy?"Terminology errors, philosophy conflicts, decision inconsistency

Workflow

code
1. READ SPEC → Understand what's proposed
       ↓
2. RUN 4 AGENTS → Parallel review
       ↓
3. SYNTHESIZE → Combine insights
       ↓
4. UPDATE SPEC → Add frontmatter, notes

Phase 1: Read Spec

Read the spec file. Skim reference docs if needed:

  • docs/definitions.md (if core concepts mentioned)
  • CLAUDE.md (for project patterns)

Phase 2: Parallel Agent Review

All 4 agents run by default. User can skip any.

code
Spec: features/p104_feature.md

═══ PERSPECTIVES (all run by default) ═══
  [x] UX - user flows, edge cases
  [x] Architect - technical feasibility, patterns, execution
  [x] Lean Coach - scope challenge, business alignment
  [x] Alignment - terminology, philosophy, strategy

Running: 4 agents

Options:
1. Run all (recommended)
2. Skip scope challenge (if confident)
3. Quick review (UX + Architect only)

Agent Files

AgentFileAlso includes thinking from
UXagents/ux.md
Architectagents/architect.mdsustainability, devils-advocate, execution-scout
Lean Coachagents/lean-coach.mdlean-startup-coach, business, alternatives
Alignmentagents/alignment.mdhypotheses, decisions, definitions, philosophy, kdd-scout

Parallel Dispatch

typescript
// All 4 agents run simultaneously
Task("UX review", { prompt: uxPrompt + specContent })
Task("Architect review", { prompt: architectPrompt + specContent })
Task("Lean Coach", { prompt: leanCoachPrompt + specContent })
Task("Alignment", { prompt: alignmentPrompt + specContent })

Phase 3: Synthesis

With 4 agents, synthesis is straightforward:

markdown
## Prep-Spec Review Summary

**Agents:** UX ✓, Architect ✓, Lean Coach ✓, Alignment ✓

### Blockers (must address)
- [ ] [Architect] Missing error handling for offline case
- [ ] [UX] No loading state defined

### Suggestions (consider)
- [ ] [Lean Coach] Could validate with mock data first
- [ ] [Alignment] Consider adding to hypotheses.md

### Conflicts to Resolve
- Architect wants abstraction, Lean Coach says YAGNI — discuss

### Execution Recommendation
- Similar to: {features/done/pN if applicable}
- MCP opportunities: {what can help}

### Post-Implementation
- Run /kdd to capture: {what}

Phase 4: Update Spec

After user reviews synthesis:

  1. Update spec frontmatter:
yaml
---
status: prepped
prepped_date: 2026-01-27
reviews:
  ux: passed
  architect: passed-with-notes
  lean-coach: passed
  alignment: passed
---
  1. Add "Prep Notes" section to spec if blockers/suggestions exist

  2. Offer to generate UAT:

    • "Want me to run /generate-uat for this spec?"

Skipping Perspectives

User can always skip:

  • "quick review" → UX + Architect only
  • "skip scope challenge" → skip Lean Coach
  • "skip prep" → proceed without review

Honor user's choice but note the spec remains unprepped.


Related Skills

  • /lean — Standalone scope challenge (runs Lean Coach methodology directly)
  • /innovate — Divergent thinking (explore 30 alternatives)
  • /ux — Standalone UX review
  • /dev — Execute implementation with TDD
  • /kdd — Record knowledge after implementation

Architecture Note

Standalone vs prep-spec:

PerspectiveIn prep-specStandalone
Lean CoachPart of 4-agent review/lean
UXPart of 4-agent review/ux
Dev thinkingVia Architect agent/dev
InnovationNot included (divergent)/innovate

Why Innovation is standalone only: Innovation is divergent thinking (explore many possibilities). Prep-spec is convergent (is this spec ready?). Run /innovate before writing the spec if you want to explore alternatives.